Live data from Hacker News

Why wordfreq will not be updated

github.com

541–542 of 542 posts

Re: Why wordfreq will not be updated

#541
post #530
post #512

Earlier quoted context omitted.

The most ridiculous aspects for me were the anthropomorphizing (Reminds me of that one Sam Altman interview a bit) and the use of "infinite", which both doesn't really work on vibes (as many have noted, while I'm sure chatGPT has been exposed to every word, its pattern of communication is very "regression to the mean" among them), but also is silly if taken literally, because unless we're counting like some quirky te…

Ah ok so anthropomorphizing and the phrase "infinite vocabulary" sounds impossible. I agree infinite vocabulary is a bit murky, and mathematically incorrect. If I wanted to be more mathematically correct I could say complete vocabulary, but I think that's actually a little less understandable to people. I did not mean infinite vocabulary in that it coins new words, just infinite as in very large to the point of being…

The anthropomorphizing pattern I picked up on was the whole phrase "chat with someone", because while I think LLMs are interesting and useful tools in a lot of ways, it's a pretty drastically different experience from talking to a person, and I think that a lot of the marketing of LLMs relies on people doing this and kind of "filling in the gaps" for what they imagine these models to be both doing and capable of. These differences are pretty stark and meaningful, and the strongest sign of that to me is that I've noticed there are a lot of people who don't bear this in mind and interact with the things often who are starting to exhibit what I would previously have identified as signs of significant social withdrawal, except instead of sounding like, I dunno, their favorite youtuber's political polemic or somesuch - which has by and large replaced the characteristic atrophy of verbal fluency we may have seen in a pre-internet era - their stilted speech trends more toward the professional and somehow both confident and airy tone of popular language models. I worry that this anthropomorphizing mindset may carry some negative cognitive consequences in the medium to long term, analogous but different from those of the rise of social media

As far as my job hunt goes, I'm not finding myself rejecting positions out of disappointment, but noticing that I'm often rejected by C-suite people after being technically vetted, often after having what I believe was a pleasant, positive, and often even generative conversation about what the company's plans with AI are and how I might help accomplish them. As I said before, I really do try to stay positive, and I think when putting my best foot forward, like in a job interview, I tend to succeed at that, but my experience leads me to be more negative about this moment in the industry when I am more candid, such as in this forum. I think if you're going out for work in whatever the tech bubble du jour is, the people you're talking to have really different biases and expectations from those hiring for more "boring" development jobs, in a way that makes me want to just go out for more general roles and lie low for a while, except that since I've been working on primarily ML-related projects for most of the last decade, it's also hard to convince people not in that area that I have adequate relevant experience. In this context and with bills to pay, it's hard to stay optimistic

Re: Why wordfreq will not be updated

#542
post #538

Earlier quoted context omitted.

I think they meant culture in the sense of knowledge that gets passed down from one generation to the next. Not a human culture of using machines, but a machine culture of using human languages.

Consider that the algorithm cannot evolve without human interaction. That's what I'm saying, it's a symbiote to us. If you consider "weights in the Instagram recommendation algorithm" to be "the machine", what we are talking about here has been happening for a long time now and has seen many generations, with each entity influencing the other. I don't think we'll have true machine culture until we have fully autonomo…

Hm, it's probably true that recommendation algorithms do something similar already, training on "human likes" that were influenced by the previous generation. But "human language" is a richer medium to carry information.

I don't think you need to be independent or autonomous to develop a culture. And a lot of human culture was passed down over generations without understanding why it worked. We just imitate the behaviour and rituals from our most successful ancestors or role models.

If new LLMs can access the past generation's knowledge of how to please human evaluators, they will use it. It's not a deliberate decision by an "agent", it's just the best text source to copy from. This is a new feedback loop between generations of assistants, and it bypasses whatever the human designer had in mind. Phrases like "it is always best to ask an expert" will pop up just because you tuned the LLM to sound like a helpful assistant, and that's what helpful assistants sound like in the training data. You'd have to actively steer the new generation away from using their ancestral knowledge.

I guess it comes down to what your definition of "culture" is. There is no targeted teaching of the next generation, for example - but is this a requirement? I agree that talking about "machine culture" right now sounds like a stretch, but now I wonder what pieces are actually missing.

Post reply on HN